PricePrice Measurement Measurement

Price measurement in the : a decade after the Boskin Report

Since the 1996 Boskin Report, BLS has made some important changes to improve the Consumer Price Index (CPI), such as the implementation of the geometric means formula to calculate basic indexes and the creation of alternative indexes to serve various user needs

David S. Johnson, he report by the U.S. Advisory Commission provides some estimate of their quantitative Stephen B. Reed, and to Study the Consumer Price Index (known impact. Kenneth J. Stewart Tmore commonly as the Boskin Report), issued on December 4, 1996, addressed the broad Conceptual basis of the CPI conceptual question of whether a cost-of-living index (COLI) should be the measurement objective Decisions about particular CPI issues are rooted of a price index and focused attention on three in the fundamental conceptual goals of the CPI. key problems inherent in the calculation of BLS remains committed to using a cost-of-living consumer price indexes: consumer substitution, index (COLI) as its theoretical goal for the CPI. quality change, and new goods. These issues The updated (online) version of the BLS received further attention in the 2002 report Handbook of Methods asserts the following: produced by an 11-member panel convened by the Committee on National Statistics entitled At As it pertains to the CPI, the COLI for the current What Price? Conceptualizing and Measuring month is based on the answer to the following Cost-of-Living and Price Indexes (known as the question: “What is the cost, at this month’s 1 market prices, of achieving the standard of CNSTAT Report). Subsequent to the Boskin living actually attained in the base period?” Report, the Bureau of Labor Statistics (BLS) This cost is a hypothetical expenditure—the reaffirmed its cost-of-living conceptual framework lowest expenditure level necessary at this and, building on prior research, introduced month’s prices to achieve the base period’s methodological changes that have addressed the living standard. The ratio of this hypothetical cost to the actual cost of the base-period substitution, quality, and new-goods issues. consumption basket in the base period is the These include the following: 1) the introduction COLI. Unfortunately, because the cost of of the geometric means formula to account for achieving a living standard cannot be observed lower-level substitution, 2) the introduction of the directly, in operational terms a COLI can only be David S. Johnson is Chained Consumer Price Index for All Urban approximated. Although the CPI cannot be said chief of the Division of to equal a cost-of-living index, the concept of Consumers (C-CPI-U) to provide an index that Housing and Household the COLI provides the CPI’s measurement Economic Statistics, accounts for upper-level substitution, 3) objective and the standard by which we define Bureau of the Census. Stephen B. Reed and expansion of the use of hedonic models to any bias in the CPI. BLS long has said that it Kenneth J. Stewart are improve the measurement of quality change, and operates within a cost-of-living framework in economists in the 4) the institution of procedures to introduce new producing the CPI. That framework has guided, Division of Consumer and will continue to guide, operational decisions Prices and Price goods into the index more quickly by more about the construction of the index.2 Indexes, Bureau of frequent updates to the item samples. This article Labor Statistics. details these methodological changes and This approach is explicitly endorsed by the

10 Monthly Labor Review May 2006 Boskin Report and supported by the CNSTAT report.3 Indeed, the time.7 Table 1 summarizes the state of estimates on CPI this framework dates to the 1961 report by the Stigler bias up through the Boskin Report.8 While there have been Committee of the National Bureau of Economic Research, substantial changes in CPI methodology since, other recent which was an important predecessor to the Boskin Report.4 studies and comments indicate a wide belief that an upward This choice arises from the way the CPI is used in practice. In bias still exists. In an update to their report, the members of the United States, the CPI is used extensively both by the Boskin commission estimated the bias at 0.8 percent as of government and private entities to make cost-of-living 1999. In a careful analysis in the Journal of Economic adjustments, which argues for a COLI framework. Such a Literature, David E. Lebow and Jeremy D. Rudd estimate the framework necessarily involves complications, because upward bias at 0.9 percent, with most of the bias coming from measuring a concept as broad as the cost of living in a new goods and quality change.9 dynamic economy and using only price data is theoretically On the other hand, some argue that the CPI understates and operationally difficult, but the cost-of-goods index (COGI) inflation, partly as a result of recent changes to address quality concept would have difficulties of its own and lead to an and substitution issues. Charles R. Hulten, for example, index that is possibly unsuitable for the way it is used.5 A argues that the quality change bias may be negative, possibly cost-of-goods index would perhaps yield different decisions so much so that it more than offsets the positive biases.10 about quality adjustments, substitution, and other CPI issues. Outside of the academic community, there is some perception The CNSTAT panel mentioned, however, that “for many that the CPI understates inflation, a view that has been (perhaps even most) purposes, the distinctions (between the articulated by some members of the finance community. COLI and COGI approaches) are less important than they Additionally, Robert Gordon, one of the members of the might seem.” Boskin commission, presented two papers (one with coauthor While the Boskin Report generated a great deal of Todd vanGoethem) at conferences in 2004 that suggested a attention, it is best understood as one chapter in a long historical downward bias in the CPI due to the treatment of history of evaluation of the conceptual foundations and rent and apparel.11 methodologies of the CPI and price indexes in general, a history chronicled in a 2005 article by Marshall Reinsdorf Substitution bias and Jack E. Triplett.6 This history includes both the development of price index theory and occasional reviews of Substitution bias arises in a fixed-weight CPI if consumers the CPI, specifically. change their purchasing behavior in response to relative price changes. Operationally, this substitution bias can be divided Bias in the CPI into upper-level substitution bias and lower-level substitution bias. This distinction between upper- and lower- The Boskin Report asserted an upward bias in the CPI of 1.1 level bias corresponds to the two-stage process involved in percent, arising mostly from biases related to substitution, calculating the CPI. The CPI is broken down into 211 item new goods, and quality change. This estimate was consistent categories and 38 areas, which are cities or groups of cities with widespread perception and with most other estimates at where prices are collected. This classification forms a matrix

Table 1. Pre-Boskin Report estimates of bias in the U.S. Consumer Price Index

Author(s) Point Estimate Interval Estimate

Advisory Commission to Study the CPI (interim report, 1995) ...... 1.0 0.7–2.0 Michael Boskin (1995) ...... 1.5 1.0–2.0 Congressional Budget Office (1995) ...... – .2–0.8 Michael R. Darby (1995) ...... 1.5 .5–2.5 W. Erwin Diewert (1995) ...... – .7–1.3 Robert J. Gordon (1995) ...... 1.7 – (1995) ...... – .5–1.5 Zvi Griliches (1995) ...... 1.0 .4–1.6 Dale W. Jorgenson (1995) ...... 1.0 .5–1.5 Jim Klumpner (1996) ...... – .3–0.5 Lebow, Roberts, and Stockton (1994) ...... – .4–1.5 Ariel Pakes (1995) ...... 0.8 – Shapiro and Wilcox (1996) ...... 1.1 .7–1.6 Wynne and Sigalla (1994) ...... less than 1.0 –

NOTE: This table is adopted from Brent R. Moulton, “Bias in the Consumer Price Index: What Is the Evidence?” Journal of Economic Perspectives, Fall 1996. Dashes indicate the estimate was not produced in that study.

Monthly Labor Review May 2006 11 Price Measurement

of 8,018 cells. Basic indexes are calculated for each cell; this is final version of the C-CPI-U is based on actual consumer the first stage of calculation, and substitution within each of behavior, rather than on assumptions about substitution these 8,018 cells is characterized as lower-level substitution. behavior. However, since expenditure data are available only The basic indexes are then aggregated into composite indexes, with a time lag, a geometric means formula is used to estimate culminating with the aggregate of all basic indexes, the All the indexes initially and then the figures are revised when the Items index. Substitution among the 8,018 different cells must final expenditure data are available. Note that the C-CPI-U is a be addressed differently and is termed upper-level distinct index from the standard CPI-U, rather than a change substitution. Substituting between Swiss and cheddar cheese to CPI-U methodology. However, the C-CPI-U provides an by consumers in a CPI area would be an example of lower- approximation of the quantitative impact of upper-level level substitution, because these items would be in the same substitution. The effect on the CPI of the change to the cell. Beef and chicken, on the other hand, are in different cells, geometric mean formula and the difference between the CPI- so substituting between them would be an example of upper- U and the C-CPI-U can be used as measures of the effects of level substitution. lower- and upper-level substitution, respectively. Table 2 Until 1999, the CPI was a modified Laspeyres index and summarizes those effects.17 used a modified Laspeyres or “Lowe” formula for both The far right columns show estimates of the effects of creating the basic indexes and aggregating to upper-level lower-level substitution, upper-level substitution, and total indexes. This formula effectively assumes zero substitution, substitution. The lower-level substitution estimate is derived as the initial quantities used in the formula are assumed to by comparing the standard CPI-U, which now uses a stay fixed after their introduction until the next expenditure geometric means formula to calculate most basic indexes, with weight update. That is, the modified Laspeyres formula the CPI-U-XL, an experimental index that has retained a assumes an elasticity of substitution of zero. It is well known Laspeyres formula at both levels of aggregation.18 The that a Laspeyres index is an upper bound to a cost of living upper-level substitution estimate is the difference between index. To the extent that consumers can and do change their the CPI-U and the C-CPI-U. All data are from December 1999 purchasing behavior in response to relative price changes, a through December 2004. Laspeyres formula will result in an upward bias in the index Table 2 also shows differences broken down by major and overstate the cost of living.12 group and for special categories. Many of the results are The Boskin Report estimated biases for both upper- and intuitive. For example, apparel, a group of goods among lower-level substitution: 0.15 percent per year for upper-level which substitution is relatively easy for consumers, has a bias and 0.25 percent per year for lower-level bias, which large lower-level effect. Because substitution may be difficult resulted in a total bias of 0.4 percent. Substitution bias had in housing and medical care, both have many strata that are been a concern in the CPI and other price indexes even before still Laspeyres and thus the lower-level effects are small. the Boskin Report,13 and BLS has addressed both upper-level Commodities have a larger effect than services. For lower- and lower-level bias in the index in the years since the report. level bias, many of these results correspond closely to earlier In 1999, the CPI converted to a geometric means formula BLS estimates. Before the change to geometric means was for item strata within which substitution is realistic, about 61 instituted—indeed before the Boskin Report was issued— percent of the index.14 The strata that remained Laspeyres are BLS created an experimental measure called the CPI-U-XG to mostly from housing and medical care; excluding rent and study the effects of a possible change to geometric means. owners’ equivalent rent, only one-seventh of the weight in The difference between the CPI-U and the CPI-U-XG averaged the CPI still uses a Laspeyres formula to calculate basic 0.27 percent per year from December 1994 to December 1996.19 indexes. The geometric means formula effectively assumes However, caution must be used in interpreting these constant relative expenditure on a given item, rather than numbers as definitive measures of lower-level substitution. constant quantity; as the relative price increases, the assumed The estimated effect of lower-level substitution (and, hence, quantity proportionally decreases. This formula thus implicitly total substitution) shown in Table 2 is slightly overstated, assumes a unitary elasticity of substitution. This geometric due to a formula bias inherent in the experimental CPI-U-XL. means formula is used in averaging of prices to create basic As background, in 1995 and 1996, BLS introduced seasoned indexes, but not in the aggregation of those indexes; hence, it samples into the CPI to eliminate a functional form bias.20 addresses only lower-level bias.15 With the adoption of the geometric means formula for most In 2002, the CPI started producing an additional index, the components of the CPI in 1999, seasoning of most samples Chained Consumer Price Index for All Urban Consumers became unnecessary and was discontinued for items using (C-CPI-U).16 This index uses a Tornqvist formula and the geometric means formula. However, the discontinuation expenditure data from both the base and current period in the of seasoned samples means that the experimental index, the upper-level aggregation to calculate the indexes. Thus, the CPI-U-XL, is upwardly biased. The effect of this bias differs

12 Monthly Labor Review May 2006 Table 2. Estimate of lower-level and upper-level substitution

[Annualized percent changes, December 1999 to December 20041]

Item CPI–U–XL CPI–U C–CPI–U Lower Upper Total

All items ...... 2.77 2.49 2.09 0.28 0.40 0.68

CPI major groups: Food and beverages ...... 2.9 2.6 2.3 .3 .3 .6 Housing ...... 3.0 3.0 2.8 .0 .2 .2 Apparel ...... –.3 –1.8 –2.2 1.5 .4 1.9 Transportation ...... 2.4 2.1 1.8 .3 .3 .6 Medical care ...... 4.5 4.4 4.3 .1 .1 .2 Recreation ...... 1.8 1.2 .7 .6 .5 1.1 Education and communication ...... 2.5 1.9 .0 .6 1.9 2.5 Education ...... 6.5 6.3 6.5 .2 –.2 .0 Communication ...... –1.4 –2.3 –4.8 .9 2.5 3.4 Other goods and services ...... 3.5 3.2 2.8 .3 .4 .7 Special aggregates: Food ...... 2.9 2.6 2.3 .3 .3 .6 Energy ...... 6.8 6.5 6.1 .3 .4 .7 All items less food and energy ...... 2.4 2.1 1.7 .3 .4 .7 Commodities and services: Commodities ...... 1.8 1.3 .6 .5 .7 1.2 Services ...... 3.5 3.3 3.2 .2 .1 .3

1 Data for the 2004 C–CPI–U are based on interim indexes. by item category, but at the All Items level, the CPI-U-XL is effect was a result of unusually large price dispersion and the perhaps 0.1 percent higher per year than its target measure, a weights for the CPI-U being relatively old at that time.24 From Laspeyres CPI. Said another way, the lower-level (and total) 2001 to 2004, the effect is fairly consistent and more modest, substitution effect columns shown in the table contain both a about 0.3 percent. substitution effect and a functional form effect, with the It should again be emphasized that while the geometric functional form effect being about 0.1 percent per year at the means formula used to address lower-level substitution is in All Items level. the standard CPI-U, upper-level substitution bias is not Moreover, while the geometric means formula corrects for addressed in the current CPI-U, but only in the C-CPI-U. Thus, the lower-level substitution bias, recent BLS research suggests if one believes that conceptually a COLI should account for that this estimate introduces a detectable upward bias in small upper-level substitution, then the C-CPI-U may represent a samples. The bias occurs because a geometric mean of sample preferable measure. of price changes will overestimate, on average, the geometric mean of all price changes in the population. As the sample New goods and quality change bias increases, the upward bias is reduced. BLS research has indicated that finite samples in CPI basic cells could yield an Perhaps the most fundamental problem in creating a price index upward bias in the estimator of 0.1 percentage point or more is that the market basket available to consumers constantly per year.21 In fact, this small sample bias could decrease the changes, and a price index must have methods in place to estimate of the upper-level substitution bias shown account for these changes. In the CPI, new goods can enter the previously.22 CPI sample in one of three ways. First, during repricing, if a Upper-level bias, as measured by the difference between the sampled item is no longer available in the sampled outlet, the CPI-U and C-CPI-U, at first appears to be larger than expected, but data collector “substitutes” to the most comparable item still these figures should be interpreted cautiously as well. BLS remaining in that outlet and begins pricing it. This is typically simulations before the C-CPI-U was first published resulted in an what is referred to as the “quality change” issue. Second, new initial estimate of 0.15 percent, later revised to 0.22 percent. As goods can also enter the CPI sample through sample rotation. seen previously, the average annual difference from 1999 to 2004 Finally, there are some new goods that do not fit neatly into the has been 0.4 percent. Chart 1 provides a graphic representation existing CPI structure (for example, cell phones before 1998).25 of the CPI-U, C-CPI-U, and C-CPI-XL from December 1999 to These goods are introduced into the CPI only during a revision December 2004. But it also is useful to examine the actual data for of the item structure. Therefore, the CPI must deal both with each year, as shown in table 3. From the table, for example, it is making quality adjustments to items discontinued in the CPI evident that the high upper-level figure is largely the result of the sample and incorporating new items into the index as quickly as anomalously high 0.80-percent effect in 2000.23 The size of this feasible.

Monthly Labor Review May 2006 13 Price Measurement

Chart 1. Consumer Price Index measures, 1999–2005 Index Index 120 120

115 115 CPI-U-XL 110 CPI-U 110 C-CPI-U 105 105

100 100

95 95

90 90 Dec. Jun. Dec. Jun. Dec. Jun. Dec. Jun. Dec. Jun. Dec. 1999 2000 2000 2001 2001 2002 2002 2003 2003 2004 2004

NOTE: December 1999 = 100. CPI-U-XL = experimental index; uses Laspeyres at both levels. CPI-U = official index; uses geometric means for averaging most prices and uses a Laspeyres framework for averaging indexes across items and areas. C-CPI-U = official index, but differs from the CPI-U in that it addresses consumer substitution across items and areas.

Table 3. All items percent changes, CPI-U-XL, CPI-U, C-CPI-U, 2000–2004

Year CPI–U–XL CPI–U C–CPI–U Lower Upper Total

2000 ...... 3.70 3.40 2.60 .30 .80 1.10 2001 ...... 1.83 1.55 1.27 .28 .28 .56 2002 ...... 2.65 2.38 2.02 .27 .36 .63 2003 ...... 2.12 1.86 1.69 .26 .17 .43 2004 ...... 3.52 3.29 2.881 .23 .41 .64

1 Data for the 2004 C-CPI-U are based on interim indexes.

Because the CPI seeks to approximate a COLI, conceptually asserted an upward bias of 0.6 percent, larger than the upper- the goal is for the CPI to be a constant-quality index. Thus, and lower-level substitution bias combined. Indeed the report when the quality of goods and services in the market basket went through a category-by-category analysis of quality changes, it is inevitable that the CPI must make some estimate bias, although some of its estimates seemed to be conjectural. of the quantitative value of such changes. This has been, and Robert Gordon conceded that this estimate might have been surely will continue to be, a source of disagreement and too high by one-tenth of 1 percent.26 In addition, David E. controversy in the CPI, because consumers have widely Lebow and Jeremy D. Rudd assert a smaller upward bias of varying preferences and, consequently, disagreement over 0.3 percent, about half of that coming from the medical care the valuation of changes in goods and services is to be category.27 expected. Arguments about a quality bias in the CPI have More recently, some have argued that there may actually come from both sides and with different levels of be a downward quality change bias. Hulten, for example, sophistication. It is widely perceived, or perhaps was widely estimates a downward bias of 0.71 percent to 0.97 percent per perceived, that much quality change goes undetected, year due to what he terms “link bias” and “quality cost bias.”28 resulting in an upward bias in the CPI. The Boskin Report, Link bias refers to a potential downward bias that could occur while noting BLS efforts to account for quality change, if manufacturers time real price increases to coincide with the

14 Monthly Labor Review May 2006 introduction of new models or redesigned products. Quality question. Often these values are estimated with hedonic cost bias is the bias that would result if the assumption of models, a technique referred to simply as hedonics. Hedonics cost elasticity of 1, implicit in CPI quality adjustment is widely considered the most promising technique for direct procedures, does not hold. Bart Hobijn also argues for at quality adjustment,33 and the CPI employs it for an increasing least the possibility of a downward bias.29 Comments by Bill number of categories of goods. In practice, the hedonic Gross are indicative of a somewhat widespread belief that approach gives BLS analysts another tool to consider when new BLS methods have a downward bias.30 BLS has maintained confronted with the problem of quality change. Table 4 that the evidence on quality bias and its direction are much summarizes the implementation of hedonic methods in the less clear than for substitution bias.31 CPI. The hedonics in the housing categories are small It is certainly true that the CPI—and indeed virtually any adjustments based on the aging of the housing units sampled. price index—faces difficulties both conceptually and The remaining hedonic adjustments are for categories that operationally in dealing with quality change, and economists together make up a fairly modest portion of the total weight of different persuasions have disagreed, and will continue to in the CPI—about 3 percent. disagree, on the merits of different approaches. However, it It is clear that hedonics has become an important tool for should be emphasized that any price index must somehow dealing with quality change in certain categories, increasing deal with quality change, so the problem amounts to a choice the ability to make direct quality adjustments. In 1999, for among different methods. The CPI has several methods it VCR’s and DVD players, imputation was used 267 times and uses to address quality change under different cir- direct quality adjustment only once. In 2001, after a hedonic cumstances, and shifts from one method to another for model was developed for this category, imputation was used several types of goods, while conceptually important, seem 92 times and direct quality adjustment 260 times. In major to have very minor quantitative impact on the All Items index. appliances, imputation was used 80 times in 1999, with no Operationally, the CPI deals with quality change in several direct quality adjustment; in 2001, imputation was used 40 different ways. For any given item being priced, the CPI times, while 80 quotes were directly quality adjusted. economic assistant in the field must determine if the item has While hedonics is an important technique for particular changed in any way—that is, if it has been replaced with a categories, it is important to emphasize that it is used for only new version. If the original and new versions are essentially a small part of the total index. Moreover, research from the the same, a commodity expert may deem them directly CPI-U Research Series (CPI-U-RS) shows that its impact on comparable and use the price comparison as if no quality indexes often has been modest and of uncertain direction. change had occurred. If the versions are substantially The CPI-U-RS was created to provide a methodologically different, then some sort of quality adjustment procedures consistent index; to this end estimates were made of the must be used. These procedures can be categorized either as quantitative index of methodological changes in the CPI since imputation or as direct quality adjustment. 1978.34 These included changes to quality adjustment There are two distinct methods of imputation in the CPI. procedures. The estimates in the research series are taken Cell-relative imputation, sometimes called “linking,” imputes from simulations described in the research for each item the price change for the noncomparable versions by the price category for which hedonics was implemented. change of all the other similar items in the same geographic In table 5, a negative sign indicates that the change to area. Thus, the price change for that quote is estimated as hedonic adjustment has caused the index to rise more slowly being the same percentage change as the price change for the (or decline more rapidly) than it would have if previous quality cell for that item stratum and index area. Class-mean adjustment procedures had been used. The inconsistency of imputation is used when price change is closely associated the effect is exemplified by the fact that the impacts for with the introduction of new lines or models, such as in the washers and dryers have the opposite sign. While the switch new vehicles category. With class-mean imputation, the price to hedonic adjustment had a significant effect on several of change is estimated from the other observations going the individual item categories, it is important to note that the through replacement at the same time that they were either net effect on the All Items index was negligible. This is quality adjusted directly or judged directly comparable.32 because the direction of these effects varied and the items in question had such a small weight. (The total relative Quality adjustment and hedonics importance of items for which hedonics have been implemented since 1998 is less than 1 percent.) Indeed, the Direct quality adjustment refers to the analyst making an net effect of hedonics from 1999 onward (which excludes estimate of the quantitative value of a quality change. This is personal computers, but includes televisions and all later done either based on manufacturer cost data or on estimates categories) on the All Items index is estimated to be less than of the value to consumers of particular features of the good in 1-hundredth of 1 percent per year, specifically +0.005 percent.

Monthly Labor Review May 2006 15 Price Measurement

Table 4. Hedonics in the CPI by date of introduction procedures. An internal BLS study looked at the effects of hedonics compared with completely omitting any quotes Date Item Weight1 where there was a quality change for the video and audio equipment categories (basically a matched model approach).36 This study used data from December 2002 to February 19882 ...... Rent 6.133 Owners’ equivalent rent 23.158 February 2005. (See table 6.) As would be expected, this January 1991 ...... Apparel3 2.160 produces quantitative differences more substantial than the January 1998 Computers4 .192 previous methods, but only for televisions and computers is January 1999 ...... Televisions .132 the effect relatively large. Table 6 also illustrates the January 2000 ...... Audio equipment (12 items) .104 differences due to quality adjustments for women’s dresses Video cameras5 .043 and computers. As suggested earlier, the quality adjustment April 2000 ...... VCR’s and DVD players5 .043 ...... used for women’s dresses causes the index to rise (or fall less July 2000 ...... Refrigerators, freezers, and rapidly). Quality adjustments for computers, however, have microwave ovens6 .165 the same effect on the index as for televisions—these College textbooks .217 adjustments cause the index to fall more rapidly.

October 2000 Washers and dryers6 .165 In the past few years, BLS has moved away from using hedonics to value the quality changes resulting from Total ...... 32.304 substitutions in computers. From January 1998 to September Total excluding housing . 3.013 Total excluding housing 2003, the CPI program used hedonic regressions, developed and apparel ...... 853 in a cooperative effort with the other price programs, as a basis to determine appropriate quality adjustment amounts 1 “Weight” represents the relative importance of components in the Consumer Price Indexes: U.S. city average, December 2004; available on for personal computers. Due to the rapid and constant the Internet at ftp://ftp.bls.gov/pub/special.requests/cpi/cpiri_2004.txt. change in PC configurations, the CPI began to move towards 2 Age-bias adjustments were introduced in 1988, and structural change an approach that uses attribute values available on the adjustments were introduced in 1989. Internet as a basis for determining the appropriate quality 3 The figure for apparel represents the total relative importance of the portions of the apparel category for which hedonic adjustment is used. adjustment amounts for personal computers. By September 4 Computer quality adjustment is now done using an attribute pricing 2003, a process of attribute cost adjustment was fully approach that uses specific manufacturer’s cost information to estimate implemented. The attribute cost adjustment process has a values for features of the good. database of 250 to 300 variables/items that are updated 5 Video cameras, VCR’s, and DVD players make up approximately 84 percent of the CPI sample for the stratum video products other than monthly. This alternative method for quality adjustments televisions. The relative importance given here is for the entire stratum. allows for more adjustments to be calculated, because many 6 Refrigerators, home freezers, microwave ovens, washers, and dryers of the items that change in a PC are not specifically covered make up approximately 75 percent of the major appliances stratum.The relative importance given here is for the entire stratum. in a hedonic model. A recent BLS study compared the quality adjustments arising This contrasts sharply with the perception that the recent from the current attribute method with those that would arise increased use of hedonics has had a substantial downward effect on the index. Table 5. Estimated impact of hedonic quality adjustment To the extent that hedonic methods were used prior to versus previous method for CPI categories in which hedonics has been introduced since 1998 1998, they tended to make the CPI slightly higher. The CPI implemented hedonic methods for many apparel categories Item Yearly effect (in percent) in 1991. Apparel is different from technology goods, in that changes in quality are not as likely to be consistently positive. Computers ...... –3.811 The estimated net affect on apparel of using hedonic Televisions ...... –.11 Audio equipment ...... 1.52 adjustment is positive; the hedonic methods make the VCR’s ...... 1.89 relevant apparel indexes higher by an estimated +0.39 percent Camcorders ...... 15 Refrigerators ...... 02 per year, compared with previous methods. In 1988, the CPI Washers ...... –.78 implemented a hedonic approach to quality adjust housing Dryers ...... 06 Microwaves ...... –.17 for the aging of the housing stock. This adjustment is College textbooks ...... –2.53 estimated to have an effect of +0.31 percent per year on the 35 affected indexes. 1 This effect was for the expenditure category Information Processing Note that these figures are estimates of the effect of Equipment, of which computers was a portion. The effect on computers switching to hedonic methods from other quality adjustment alone was about –6.5 percent.

16 Monthly Labor Review May 2006 Table 6. Annualized percentages with and without quality adjustment Annualized percent Item Annualized percent change Difference change without substitution1

Video and audio products (Dec. 2002 – Feb. 2005) ...... 0.46 0.68 –0.22 Televisions ...... –12.86 –10.92 –1.94 Video products other than televisions ...... –13.52 –13.43 –.10 Audio products ...... –6.66 -6.26 –.41 Women’s dresses (Dec. 2002 – Feb. 2005) ...... –3.83 –3.99 +.16 Computers and peripheral equipment (Mar. 2004 – Sep. 2004)...... –9.78 –6.60 –3.18

1 Because the phrase “without substitution” is vague, the following briefly describes the actual procedures used in this simulation: for video and audio products, the simulation removed quality adjustments and imputed the price change for those quotes, while directly compared substitutions were left alone. For women’s dresses, the simulation removed quality adjustments and either directly compared or imputed the price change for those quotes, while the directly compared substitutions were left alone. For personal computers, the simulation removed all substitutions regardless of whether actual quality change was present. from the hedonic method.37 This study compared 6 months of percent of the outlet sample each year to 25 percent, so that adjustments (April 2004 to September 2004). Compared with the the entire sample is rotated every 4 years instead of every 5 original hedonic method, the new attribute method results in a years. Moreover, some items in selected categories that tend slightly larger decline in the index (an annualized rate of –9.78 to change rapidly are rotated every 2 years. Thus, the market percent, compared with –8.58 percent for the hedonic method). basket of the CPI is considerably more up to date than it used to be, particularly in terms of high-tech goods. For example, Updating the market basket the new procedures have resulted in greater representation in the sample for technology items like flat panel televisions Along with quality adjustment of goods in the sample, there and digital video recorders. is the issue of goods entering the economy that are not in the One final recommendation of the Boskin Report was for sample. Given the COLI concept, it is crucial that the CPI get BLS to improve its mechanism for bringing in outside new goods into the sample quickly, in order to have a market information, research, and expertise. In 2000, the Federal basket that accurately reflects consumer purchases. Economic Statistics Advisory Committee (FESAC) was Additionally, it allows the CPI to capture some of the created.This group meets periodically with BLS repre- consumer surplus when new goods enter the economy and sentatives and provides a nexus between BLS and the decline steadily in price, as sometimes happens with new academic research community. FESAC has allowed the CPI to technology goods; failure to capture this surplus has been exchange ideas with the academic and research community seen as a possible source of upward bias in the CPI. more efficiently, acting as a tool for the CPI both to transmit Although BLS has chosen not to attempt to reflect its latest research and methodology to the academic sector consumer surplus in the index explicitly—consistent with the and to receive new research relevant to the CPI. recommendation of the CNSTAT panel—the CPI program has taken several steps in recent years to keep the market basket THE BOSKIN REPORT FOCUSED ATTENTION ON THE CPI and some up to date. Since 2002, updated expenditure weights based particular sources of possible bias in the index, but improving upon consumer expenditure surveys have been introduced the CPI is an ongoing process. Since the 1996 report, there every 2 years (as opposed to roughly every 10 years in the have been important changes to improve the index. The past). Moreover, the lag time from survey to implementation implementation of the geometric means formula to calculate is shorter, and the survey is completed in a shorter time. The basic indexes addressed lower-level substitution bias, and result is that weights used in the CPI reflect much more recent the creation of the C-CPI-U provides a measure that accounts consumer behavior than in years past. This probably results for upper-level substitution. While the case for quality change in a smaller increase in the index; for 2004, the increase in the bias is much less clear cut, the expanded use of hedonic index was 0.06 percent lower than it would have been had the models to adjust directly for changes in quality has given old weights been in place. This figure is consistent with BLS analysts another sophisticated option with which to estimates of the impact of past revisions, which have usually, address this issue. However, compared with the new methods though not always, indicated that the weight update tends to used to address consumer substitution, and in contrast to cause a slightly lower rate of growth in the index. widespread perception, these changes have not had an Additionally, the CPI has changed its sample and outlet important quantitative effect on the All Items index. More rotation procedures. In 1998, the CPI went from rotating 20 frequent weight updates and sample rotation mean that the

Monthly Labor Review May 2006 17 Price Measurement

market basket used in calculating the CPI is more up to date and The CPI will continue to evaluate and improve its methodologies reflective of current consumer behavior than it ever has been. in order to produce the most accurate index possible.

Notes

1 Charles L. Schultze and Christopher Mackie, eds., At What Price? 16 For a detailed discussion of this index, see Robert Cage, John S. Conceptualizing and Measuring Cost-of-Living and Price Indexes, Panel Greenlees, and Patrick Jackman, “Introducing the Chained Consumer on Conceptual, Measurement, and Other Statistical Issues in Developing Price Index.” Paper presented at Seventh Meeting of the International Cost-of-Living Indexes (Washington, National Academy Press, 2002). Working Group on Price Indices, Paris, France, May 2003.

2 “Chapter 17, The Consumer Price Index,” BLS Handbook of 17 Unlike the CPI-U and CPI-U-XL, major group and other subaggregate Methods, updated online version (Bureau of Labor Statistics, April 10, C-CPI-U indexes are not independent. For example, the Food and 2006), pp. 2–3; available on the Internet at http://www.bls.gov/ Beverage C-CPI-U index reflects average price change among food and opub/hom/pdf/homch17.pdf. For more information on the beverage items, but because relative price change for items in other historical context of the COLI approach used in the CPI, see John S. major groups (Housing, Transportation, for example) may affect the Greenlees, “A Bureau of Labor Statistics Perspective on Bias in the level of expenditure on food items, the Food and Beverage C-CPI-U Consumer Price Index,” Federal Reserve Bank of St. Louis Review, index is conditional upon price change in other major groups. May 1997, pp. 175–78. Moreover, C-CPI-U indexes are not precisely consistent in aggregation, although in practice they are very close. That is, an expenditure 3 For a detailed discussion of the conceptual foundations of price weighted average of the subindexes may not yield the exact estimate indexes, see Schultze and Mackie, eds., At What Price? Conceptualizing of All Items price change as the official All Items index. Additionally, and Measuring Cost-of-Living and Price Indexes, pp. 43–73. The the difference between the C-CPI-U and published CPI-U is being used as CNSTAT report approves of a COLI approach, but does not explicitly a measure of upper-level substitution even though the published CPI-U recommend it to the exclusion of other approaches. is technically a Lowe index. Thus, while this table does give an idea of 4 For a detailed history of the Consumer Price Index, see Marshall the relative magnitude of the substitution effects, it should be Reinsdorf and Jack E. Triplett, “A Review of Reviews: Ninety Years interpreted with caution for the major groups and other subaggregates. of Professional Thinking About the Consumer Price Index.” Paper Note, however, that BLS research on upper-level substitution bias, presented at the CRIW Conference on Price Index Concepts and which had preceded the development of the C-CPI-U and had been Measurement, Vancouver, Canada, June 28–29, 2004. provided to the Boskin Commission, was based on the comparison of 5 a true Laspeyres index with Fisher and Tornqvist superlative indexes. For a discussion of the COLI and COGI debate, see Reinsdorf and Triplett, “A Review of Reviews,” pp. 34–41. 18 The CPI-U-XL is an unofficial index that uses a Laspeyres formula 6 Ibid. to average the prices within basic item area cells, as well as aggregating those indexes. 7 See Brent R. Moulton, “Bias in the Consumer Price Index: What 19 is the Evidence?” Journal of Economic Perspectives, Fall 1996. Data from which this figure is derived are available on the BLS website at http://www.bls.gov/cpi/cpigmatab.htm. 8 The table is adopted from Moulton; ibid. 20 See Ken Stewart, “Improving Sample Rotation Procedures,” CPI 9 David E. Lebow and Jeremy D. Rudd, “Measurement Error in the Detailed Report, October 1994, pp.7–8; and “Extending the Consumer Price Index: Where Do We Stand?” Journal of Economic Improvements in CPI Sample Rotation Procedures and Improving the Literature, March 2003, 159–201. Procedures for Substitute Items,” CPI Detailed Report, March, 1996, 10 Charles R. Hulten, “Quality Change in the CPI,” Federal Reserve pp.4–5. Bank of St. Louis Review, May 2001, pp. 87–111. 21 See Robert McClelland and Marshall Reinsdorf, “Small Sample Bias 11 See Robert Gordon, “Apparel Prices 1914–93 and the Hulten/ in Geometric Mean and Seasoned CPI Component Indexes,” Bureau of Bruegel Paradox.” Paper presented at CRIW Conference on Price Index Labor Statistics Working Paper 324, August 1999; and Ralph Bradley, Concepts and Measurement, Vancouver, Canada, 2004; and Robert “Analytical Bias Reduction for Small Samples in the US Consumer Price Gordon and Todd vanGoethem, “A Century of Housing Shelter Prices: Index,” BLS unpublished manuscript, September 3, 2004. Is There a Downward Bias in the CPI?” NBER Working Paper No. 22 See Bradley, “Analytical Bias Reduction for Small Samples in the 11776 (National Bureau of Economic Research, November 2005). US Consumer Price Index.” 12 Technically, the CPI is a Lowe price index, not a Laspeyres index. 23 Robert Cage and coauthors estimate that the size of the differenceis For more on this distinction, see Bert M. Balk and W. Erwin Diewert, only 0.6 percent if the biennial weights are used in the construction of “The Lowe Consumer Price Index and its Substitution Bias,” the CPI-U. In addition, this large difference is also apparent in the difference Discussion Paper 04–07 (Department of , University of between the PCE chain-weight and fixed-weight indexes for 2000. British Columbia), July 2004. In general, the modified Laspeyres or 24 Lowe index does not have the property of being an upper bound to a Weights for the final version of the C-CPI-U are from cost-of-living index. This follows because the comparisons of contemporaneous expenditure data and so the CPI-U weights are always, intermediate values of a Laspeyres index subsequent to the base period in a sense, older than those of the C-CPI-U. However, since the CPI-U do not have this property. weights are updated every 2 years, the time gap between the weights of the two indexes varies over time. 13 See Ana M. Aizcorbe and Patrick C. Jackman, “The Commodity 25 Substitution Effect in CPI Data, 1982–1991: Anatomy of a Price For more on these discussions, see Schultze and Mackie, eds., At Change,” Monthly Labor Review, December 1993, pp. 25–33. What Price? Conceptualizing and Measuring Cost-of-Living and Price Indexes, p. 31. 14 See “Chapter 17, The Consumer Price Index,” BLS Handbook of Methods, for item categories that use the Laspeyres formula. 26 See Robert Gordon, “Apparel Prices 1914–93 and the Hulten/ Bruegel Paradox.” 15 Reinsdorf and Triplett suggest other reasons beyond substitution to use the geometric mean estimator. 27 David E. Lebow and Jeremy D. Rudd, “Measurement Error in the

18 Monthly Labor Review May 2006 Consumer Price Index: Where Do We Stand?” Journal of Economic 32 This discussion follows that in the BLS Handbook of Methods. Literature, March 2003, pp. 159–201. 33 See, Schultze and Mackie, eds., At What Price? Conceptualizing 28 Hulten, “Quality Change in the CPI.” and Measuring Cost-of-Living and Price Indexes, p. 122. 29 Bart Hobijn, “On Both Sides of the Quality Bias in Price Indexes,” 34 Kenneth J. Stewart and Stephen B. Reed, “Consumer Price Index Staff Report 157 (Federal Reserve Bank of , 2002); on the Using Current Methods,” Monthly Labor Review, June 1999, pp. 29–38. Internet at http://ideas.repec.org/p/fip/fednsr/157.html#provider. 35 In the past 3 years, the average adjustment has fallen to 0.26 30 Bill Gross, “Haute Con Job,” Investment Outlook (PIMCO Bonds, percent per year. October 2004). 36 David S. Johnson and Craig Brown, Internal BLS memorandum, 31 This is made especially clear in Brent R. Moulton, “Bias in the April 2005. Consumer Price Index: What is the Evidence?”; and in John S. Greenlees, “A Bureau of Labor Statistics Perspective on Bias in the 37 David S. Johnson and Joe Chelena, Internal BLS memorandum, Consumer Price Index.” February 2005.

Monthly Labor Review May 2006 19